This repository contains code for training a GFlowNet encoder-decoder transformer to sample trajectories from the sliding puzzle game. To start training, run

python main.py. The board size can be modified using the --boardwidth argument. Training hyperparameters can also be modified. Their descriptions are found in main.py.

The reward function is $e^{-n}$, where $n$ is the number of mismatching squares (Hamming distance) between the final board and the solved board.